What are long-term issues involved in machine learning? — - technical debt - lack of clear abstraction barriers - changing anything changes everything - feedback loop - usage based on your model changes the model - attractive nuisance (using a successful model in one domain where it doesn't fit in another) - non-stationarity - stick with current data, or get new, and how much of old data to reuse - tracking data dependencies - where did the data come from, how to get new data